Plugins vs Agent Skills
A Skill is one capability; a plugin is the box you ship several of them in. The question is not which to use, but whether you are building or distributing.
Last verified August 21, 2026
The short answer
Write a Skill when you are adding a capability. Publish a plugin when you are distributing one or more capabilities to other people — a plugin can contain skills, commands, hooks and MCP server configuration in a single installable unit.
Can you use both? Yes — most plugins exist precisely to distribute skills alongside the other pieces they need.
Side by side
Option A
Agent Skills
A folder of instructions, scripts and resources, loaded on demand. Works in the Claude apps, Claude Code, the API and the Agent SDK.
Pick it when
You are solving your own problem, or the capability stands alone and needs nothing installed alongside it.
Option B
Plugins
An installable bundle distributed through marketplaces. Can carry skills, slash commands, hooks, MCP servers and subagent definitions together.
Pick it when
Other people need to install it, or the capability needs more than instructions — a hook to enforce something, an MCP server to reach something.
Containment, not competition
A plugin contains skills. They are not alternatives, and asking “should I use a plugin or a skill” is like asking whether to use a file or a zip archive.
| Skill | Plugin | |
|---|---|---|
| What it is | One capability | A distribution bundle |
| Can contain | Instructions, scripts, resources | Skills, commands, hooks, MCP servers, subagents |
| Installed from | A folder, or a plugin | A marketplace |
| Portable across | Claude apps, Code, API, SDK | Claude Code, Cowork |
| Open standard | Yes | No |
Note the last two rows: Skills are the more portable artefact and are published as a cross-platform open standard. Plugins are the richer container but a narrower one.
When the bundle is genuinely necessary
A skill alone is enough when the capability is pure procedure. You need a plugin when the capability has moving parts:
- It must enforce something → a hook, which a skill cannot guarantee
- It must reach something → an MCP server configuration
- It needs a command users invoke explicitly
- It ships a subagent with constrained tools
For organisations
Anthropic documents plugin dependency version constraints, CLI plugin hints, and organisational plugin relevance. That combination says the intended use is teams standardising on a set of capabilities, not individuals collecting tools.
If you are rolling Claude Code out across a company, plugins plus server-managed settings are how you make a default rather than a suggestion.
Reviewing a community plugin
Three questions, every time:
- What hooks does it register? Those run shell commands with your privileges.
- What MCP servers does it configure? Those hold credentials and can write.
- What permissions do those two imply?
Everything community-maintained in this reference’s ecosystem directory carries that warning for this reason.
Common misconceptions
- "Plugins are the Claude Code version of Skills." No. Skills work in Claude Code too. Plugins are a packaging and distribution mechanism, not a surface.
- "Installing a plugin is low risk because it is just configuration." A plugin can register hooks that run shell commands and configure MCP servers that hold credentials. It is code you are choosing to run.
- "Plugins are Claude Code only." They also customise Cowork.
Sources
- 01Create plugins — Claude Code docsOfficialcode.claude.com
- 02Agent Skills announcementOfficialclaude.com
- 03Customize Cowork with pluginsOfficialclaude.com